Description
(3) A large number of features. Different from the data sets in traditional data mining, the features observed in manufacturing analysis
may be thousands. Care must therefore be taken to avoid that machine learning algorithms can only work with reduced datasets (i.e.
datasets with a small number of features).
(4) Multicollinearity, when products pass through the assembly line, different measurement methods are taken at different stations
in the production process. Some of these measurements can be highly correlated, however many machine learning and data mining
algorithm properties are independent of each other, and multicollinearity issues should be carefully studied for the proposed analysis method.
(5) Classification imbalance problem, where there is a huge imbalance between good and bad parts (or scrap, that is, parts that do not
pass quality control testing). Ratios may range from 9:1 to even lower than 99,000,000:1. It is difficult to distinguish good parts from scrap
using standard classification techniques, so several methods for handling class imbalance have been proposed and applied to manufacturing analysis [8].
(6) Non-stationary data, the underlying manufacturing process may change due to various factors such as changes in suppliers
or operators and calibration deviations in machines. There is therefore a need to apply more robust methods to the non-stationary
nature of the data. (7) Models can be difficult to interpret, and production and quality control engineers need to understand the analytical
solutions that inform process or design changes. Otherwise the generated recommendations and decisions may be ignored.
Reviews
There are no reviews yet.